Feature extraction is a vital pre-processing step for time series analysis. Signatures from rough path theory have been proposed as an attractive feature extraction method for machine learning in general. In this study, a pilot study on an Electrocardiogram (ECG) Human activity recognition (HAR) dataset is performed. The signatures for a path of ECG signals were calculated and then tested with different basic classifiers like random forest, and a plain neural network. It is observed that using signatures for features improves the classification accuracy from 74% for raw data to 92% using signatures as features.